Microsoft's Windows Applied Sciences Group launches Aion 1.0 Instruct, a 2.29B parameter SLM designed for high-performance on-device tasks and Windows OS integration.

On November 18, 2026, Microsoft officially released Aion 1.0 Instruct, a significant milestone in the evolution of Small Language Models (SLMs). Published open-weight by the Windows Applied Sciences Group, this model serves as the foundational intelligence for the future of the Windows ecosystem. While it was first teased at Build 2026 with a preview sample that garnered 788 downloads in just 36 hours, the full release marks its transition from an experimental tool to a production-ready engine.
This release is strategically timed to pave the way for the Windows OS SLM, which will be available at general availability by the end of January 2027. By providing the weights openly under an MIT license, Microsoft is empowering developers to build privacy-first, cloud-free applications that run directly on local hardware, bridging the gap between massive LLMs and efficient edge computing.
Aion 1.0 Instruct is a highly optimized 2.29B parameter model featuring 28 layers. It utilizes Grouped-Query Attention (GQA) with 24 query heads and 8 key-value heads to maximize inference efficiency. A standout architectural choice is its hybrid attention mechanism: it employs 21 sliding-window layers (with a 2048 window size) combined with 7 full-attention layers. This allows the model to maintain high performance while significantly reducing the computational overhead typically associated with long-context processing.
The model supports a massive 32,768-token context window, enabled by YaRN scaling (x16). This allows developers to process multi-page documents, lengthy email threads, and complex technical specifications entirely on-device. The vocabulary size is set at 200K, and the model is released in unquantized bfloat16 format to ensure maximum precision for initial deployment. Furthermore, a dedicated ONNX-WebGPU variant has been released simultaneously, allowing the model to run directly within modern web browsers.
Aion 1.0 Instruct is the direct successor to Phi Silica, but it represents a massive leap in capability. The training corpus was significantly expanded using synthetic data specifically curated for math, coding, common sense reasoning, and theory of mind. While the model remains a specialist in instruction following rather than a general-purpose world-knowledge engine, its benchmark scores reflect a highly capable reasoning agent for its size.
In rigorous testing using the lm-evaluation-harness, Aion 1.0 Instruct demonstrated impressive proficiency in logic and instruction adherence. It achieved a GSM8K score of 74.91 and a MATH-500 score of 48.00, showcasing its strength in mathematical reasoning. For instruction following, it scored 83.90 on IFEval and 64.70 on Multi-IF. Multilingual capabilities are also present, with 8 supported languages (English, German, Spanish, Portuguese, Italian, French, Chinese, and Japanese) making up approximately 6% of the training mix.
As Aion 1.0 Instruct is primarily designed for local, on-device deployment and open-weight usage, there is no centralized hosted API pricing provided by Microsoft for this specific model release. Developers are encouraged to run the model on their own infrastructure using PyTorch, ONNX, or GGUF.
For those looking to integrate it into web applications, the ONNX-WebGPU variant provides a zero-cost inference path where the user's local hardware provides the compute. This 'cloud-free' approach eliminates per-token costs for the developer and ensures maximum privacy for the end-user.
Aion 1.0 Instruct is explicitly engineered for specific tasks rather than broad, multi-step complex reasoning or general world knowledge retrieval. Its architecture is optimized for high-speed, low-latency interactions that require understanding context without needing a connection to a massive data center.
The primary use cases include document summarization, text rewriting, conversational chat interfaces, and deep document comprehension. Because it can handle a 32K context window, it is ideal for RAG (Retrieval-Augmented Generation) workflows where the retrieved context must be processed locally to maintain data privacy. It is also a perfect candidate for 'Copilot-style' features within local applications, such as drafting emails or explaining technical specifications.
Developers can access Aion 1.0 Instruct immediately through various open-source channels. The model is available in bfloat16, with quantized versions expected to follow shortly. Support is provided for PyTorch, ONNX, and GGUF, ensuring compatibility with most modern AI deployment stacks.
Note that because the model has a knowledge cutoff of December 2025, users must ensure they are using transformers 5.x and set `trust_remote_code=True` when loading via the Hugging Face ecosystem. For web developers, the ONNX-WebGPU variant is the recommended path for immediate browser integration. You can find the initial preview samples and documentation on Microsoft's GitHub repository.